1 //===-- ShadowStackGCLowering.cpp - Custom lowering for shadow-stack gc ---===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the custom lowering code required by the shadow-stack GC 11 // strategy. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/CodeGen/Passes.h" 16 #include "llvm/ADT/StringExtras.h" 17 #include "llvm/CodeGen/GCStrategy.h" 18 #include "llvm/IR/CallSite.h" 19 #include "llvm/IR/IRBuilder.h" 20 #include "llvm/IR/IntrinsicInst.h" 21 #include "llvm/IR/Module.h" 22 23 using namespace llvm; 24 25 #define DEBUG_TYPE "shadowstackgclowering" 26 27 namespace { 28 29 class ShadowStackGCLowering : public FunctionPass { 30 /// RootChain - This is the global linked-list that contains the chain of GC 31 /// roots. 32 GlobalVariable *Head; 33 34 /// StackEntryTy - Abstract type of a link in the shadow stack. 35 /// 36 StructType *StackEntryTy; 37 StructType *FrameMapTy; 38 39 /// Roots - GC roots in the current function. Each is a pair of the 40 /// intrinsic call and its corresponding alloca. 41 std::vector<std::pair<CallInst *, AllocaInst *>> Roots; 42 43 public: 44 static char ID; 45 ShadowStackGCLowering(); 46 47 bool doInitialization(Module &M) override; 48 bool runOnFunction(Function &F) override; 49 50 private: 51 bool IsNullValue(Value *V); 52 Constant *GetFrameMap(Function &F); 53 Type *GetConcreteStackEntryType(Function &F); 54 void CollectRoots(Function &F); 55 static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, 56 Value *BasePtr, int Idx1, 57 const char *Name); 58 static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, 59 Value *BasePtr, int Idx1, int Idx2, 60 const char *Name); 61 }; 62 } 63 64 INITIALIZE_PASS_BEGIN(ShadowStackGCLowering, "shadow-stack-gc-lowering", 65 "Shadow Stack GC Lowering", false, false) 66 INITIALIZE_PASS_DEPENDENCY(GCModuleInfo) 67 INITIALIZE_PASS_END(ShadowStackGCLowering, "shadow-stack-gc-lowering", 68 "Shadow Stack GC Lowering", false, false) 69 70 FunctionPass *llvm::createShadowStackGCLoweringPass() { return new ShadowStackGCLowering(); } 71 72 char ShadowStackGCLowering::ID = 0; 73 74 ShadowStackGCLowering::ShadowStackGCLowering() 75 : FunctionPass(ID), Head(nullptr), StackEntryTy(nullptr), 76 FrameMapTy(nullptr) { 77 initializeShadowStackGCLoweringPass(*PassRegistry::getPassRegistry()); 78 } 79 80 namespace { 81 /// EscapeEnumerator - This is a little algorithm to find all escape points 82 /// from a function so that "finally"-style code can be inserted. In addition 83 /// to finding the existing return and unwind instructions, it also (if 84 /// necessary) transforms any call instructions into invokes and sends them to 85 /// a landing pad. 86 /// 87 /// It's wrapped up in a state machine using the same transform C# uses for 88 /// 'yield return' enumerators, This transform allows it to be non-allocating. 89 class EscapeEnumerator { 90 Function &F; 91 const char *CleanupBBName; 92 93 // State. 94 int State; 95 Function::iterator StateBB, StateE; 96 IRBuilder<> Builder; 97 98 public: 99 EscapeEnumerator(Function &F, const char *N = "cleanup") 100 : F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {} 101 102 IRBuilder<> *Next() { 103 switch (State) { 104 default: 105 return nullptr; 106 107 case 0: 108 StateBB = F.begin(); 109 StateE = F.end(); 110 State = 1; 111 112 case 1: 113 // Find all 'return', 'resume', and 'unwind' instructions. 114 while (StateBB != StateE) { 115 BasicBlock *CurBB = StateBB++; 116 117 // Branches and invokes do not escape, only unwind, resume, and return 118 // do. 119 TerminatorInst *TI = CurBB->getTerminator(); 120 if (!isa<ReturnInst>(TI) && !isa<ResumeInst>(TI)) 121 continue; 122 123 Builder.SetInsertPoint(TI->getParent(), TI); 124 return &Builder; 125 } 126 127 State = 2; 128 129 // Find all 'call' instructions. 130 SmallVector<Instruction *, 16> Calls; 131 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 132 for (BasicBlock::iterator II = BB->begin(), EE = BB->end(); II != EE; 133 ++II) 134 if (CallInst *CI = dyn_cast<CallInst>(II)) 135 if (!CI->getCalledFunction() || 136 !CI->getCalledFunction()->getIntrinsicID()) 137 Calls.push_back(CI); 138 139 if (Calls.empty()) 140 return nullptr; 141 142 // Create a cleanup block. 143 LLVMContext &C = F.getContext(); 144 BasicBlock *CleanupBB = BasicBlock::Create(C, CleanupBBName, &F); 145 Type *ExnTy = 146 StructType::get(Type::getInt8PtrTy(C), Type::getInt32Ty(C), nullptr); 147 Constant *PersFn = F.getParent()->getOrInsertFunction( 148 "__gcc_personality_v0", FunctionType::get(Type::getInt32Ty(C), true)); 149 LandingPadInst *LPad = 150 LandingPadInst::Create(ExnTy, PersFn, 1, "cleanup.lpad", CleanupBB); 151 LPad->setCleanup(true); 152 ResumeInst *RI = ResumeInst::Create(LPad, CleanupBB); 153 154 // Transform the 'call' instructions into 'invoke's branching to the 155 // cleanup block. Go in reverse order to make prettier BB names. 156 SmallVector<Value *, 16> Args; 157 for (unsigned I = Calls.size(); I != 0;) { 158 CallInst *CI = cast<CallInst>(Calls[--I]); 159 160 // Split the basic block containing the function call. 161 BasicBlock *CallBB = CI->getParent(); 162 BasicBlock *NewBB = 163 CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont"); 164 165 // Remove the unconditional branch inserted at the end of CallBB. 166 CallBB->getInstList().pop_back(); 167 NewBB->getInstList().remove(CI); 168 169 // Create a new invoke instruction. 170 Args.clear(); 171 CallSite CS(CI); 172 Args.append(CS.arg_begin(), CS.arg_end()); 173 174 InvokeInst *II = 175 InvokeInst::Create(CI->getCalledValue(), NewBB, CleanupBB, Args, 176 CI->getName(), CallBB); 177 II->setCallingConv(CI->getCallingConv()); 178 II->setAttributes(CI->getAttributes()); 179 CI->replaceAllUsesWith(II); 180 delete CI; 181 } 182 183 Builder.SetInsertPoint(RI->getParent(), RI); 184 return &Builder; 185 } 186 } 187 }; 188 } 189 190 191 Constant *ShadowStackGCLowering::GetFrameMap(Function &F) { 192 // doInitialization creates the abstract type of this value. 193 Type *VoidPtr = Type::getInt8PtrTy(F.getContext()); 194 195 // Truncate the ShadowStackDescriptor if some metadata is null. 196 unsigned NumMeta = 0; 197 SmallVector<Constant *, 16> Metadata; 198 for (unsigned I = 0; I != Roots.size(); ++I) { 199 Constant *C = cast<Constant>(Roots[I].first->getArgOperand(1)); 200 if (!C->isNullValue()) 201 NumMeta = I + 1; 202 Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr)); 203 } 204 Metadata.resize(NumMeta); 205 206 Type *Int32Ty = Type::getInt32Ty(F.getContext()); 207 208 Constant *BaseElts[] = { 209 ConstantInt::get(Int32Ty, Roots.size(), false), 210 ConstantInt::get(Int32Ty, NumMeta, false), 211 }; 212 213 Constant *DescriptorElts[] = { 214 ConstantStruct::get(FrameMapTy, BaseElts), 215 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), Metadata)}; 216 217 Type *EltTys[] = {DescriptorElts[0]->getType(), DescriptorElts[1]->getType()}; 218 StructType *STy = StructType::create(EltTys, "gc_map." + utostr(NumMeta)); 219 220 Constant *FrameMap = ConstantStruct::get(STy, DescriptorElts); 221 222 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems 223 // that, short of multithreaded LLVM, it should be safe; all that is 224 // necessary is that a simple Module::iterator loop not be invalidated. 225 // Appending to the GlobalVariable list is safe in that sense. 226 // 227 // All of the output passes emit globals last. The ExecutionEngine 228 // explicitly supports adding globals to the module after 229 // initialization. 230 // 231 // Still, if it isn't deemed acceptable, then this transformation needs 232 // to be a ModulePass (which means it cannot be in the 'llc' pipeline 233 // (which uses a FunctionPassManager (which segfaults (not asserts) if 234 // provided a ModulePass))). 235 Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true, 236 GlobalVariable::InternalLinkage, FrameMap, 237 "__gc_" + F.getName()); 238 239 Constant *GEPIndices[2] = { 240 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0), 241 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)}; 242 return ConstantExpr::getGetElementPtr(GV, GEPIndices); 243 } 244 245 Type *ShadowStackGCLowering::GetConcreteStackEntryType(Function &F) { 246 // doInitialization creates the generic version of this type. 247 std::vector<Type *> EltTys; 248 EltTys.push_back(StackEntryTy); 249 for (size_t I = 0; I != Roots.size(); I++) 250 EltTys.push_back(Roots[I].second->getAllocatedType()); 251 252 return StructType::create(EltTys, "gc_stackentry." + F.getName().str()); 253 } 254 255 /// doInitialization - If this module uses the GC intrinsics, find them now. If 256 /// not, exit fast. 257 bool ShadowStackGCLowering::doInitialization(Module &M) { 258 bool Active = false; 259 for (Function &F : M) { 260 if (F.hasGC() && F.getGC() == std::string("shadow-stack")) { 261 Active = true; 262 break; 263 } 264 } 265 if (!Active) 266 return false; 267 268 // struct FrameMap { 269 // int32_t NumRoots; // Number of roots in stack frame. 270 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots. 271 // void *Meta[]; // May be absent for roots without metadata. 272 // }; 273 std::vector<Type *> EltTys; 274 // 32 bits is ok up to a 32GB stack frame. :) 275 EltTys.push_back(Type::getInt32Ty(M.getContext())); 276 // Specifies length of variable length array. 277 EltTys.push_back(Type::getInt32Ty(M.getContext())); 278 FrameMapTy = StructType::create(EltTys, "gc_map"); 279 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy); 280 281 // struct StackEntry { 282 // ShadowStackEntry *Next; // Caller's stack entry. 283 // FrameMap *Map; // Pointer to constant FrameMap. 284 // void *Roots[]; // Stack roots (in-place array, so we pretend). 285 // }; 286 287 StackEntryTy = StructType::create(M.getContext(), "gc_stackentry"); 288 289 EltTys.clear(); 290 EltTys.push_back(PointerType::getUnqual(StackEntryTy)); 291 EltTys.push_back(FrameMapPtrTy); 292 StackEntryTy->setBody(EltTys); 293 PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy); 294 295 // Get the root chain if it already exists. 296 Head = M.getGlobalVariable("llvm_gc_root_chain"); 297 if (!Head) { 298 // If the root chain does not exist, insert a new one with linkonce 299 // linkage! 300 Head = new GlobalVariable( 301 M, StackEntryPtrTy, false, GlobalValue::LinkOnceAnyLinkage, 302 Constant::getNullValue(StackEntryPtrTy), "llvm_gc_root_chain"); 303 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) { 304 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy)); 305 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage); 306 } 307 308 return true; 309 } 310 311 bool ShadowStackGCLowering::IsNullValue(Value *V) { 312 if (Constant *C = dyn_cast<Constant>(V)) 313 return C->isNullValue(); 314 return false; 315 } 316 317 void ShadowStackGCLowering::CollectRoots(Function &F) { 318 // FIXME: Account for original alignment. Could fragment the root array. 319 // Approach 1: Null initialize empty slots at runtime. Yuck. 320 // Approach 2: Emit a map of the array instead of just a count. 321 322 assert(Roots.empty() && "Not cleaned up?"); 323 324 SmallVector<std::pair<CallInst *, AllocaInst *>, 16> MetaRoots; 325 326 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 327 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) 328 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++)) 329 if (Function *F = CI->getCalledFunction()) 330 if (F->getIntrinsicID() == Intrinsic::gcroot) { 331 std::pair<CallInst *, AllocaInst *> Pair = std::make_pair( 332 CI, 333 cast<AllocaInst>(CI->getArgOperand(0)->stripPointerCasts())); 334 if (IsNullValue(CI->getArgOperand(1))) 335 Roots.push_back(Pair); 336 else 337 MetaRoots.push_back(Pair); 338 } 339 340 // Number roots with metadata (usually empty) at the beginning, so that the 341 // FrameMap::Meta array can be elided. 342 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end()); 343 } 344 345 GetElementPtrInst *ShadowStackGCLowering::CreateGEP(LLVMContext &Context, 346 IRBuilder<> &B, Value *BasePtr, 347 int Idx, int Idx2, 348 const char *Name) { 349 Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), 350 ConstantInt::get(Type::getInt32Ty(Context), Idx), 351 ConstantInt::get(Type::getInt32Ty(Context), Idx2)}; 352 Value *Val = B.CreateGEP(BasePtr, Indices, Name); 353 354 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 355 356 return dyn_cast<GetElementPtrInst>(Val); 357 } 358 359 GetElementPtrInst *ShadowStackGCLowering::CreateGEP(LLVMContext &Context, 360 IRBuilder<> &B, Value *BasePtr, 361 int Idx, const char *Name) { 362 Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), 363 ConstantInt::get(Type::getInt32Ty(Context), Idx)}; 364 Value *Val = B.CreateGEP(BasePtr, Indices, Name); 365 366 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 367 368 return dyn_cast<GetElementPtrInst>(Val); 369 } 370 371 /// runOnFunction - Insert code to maintain the shadow stack. 372 bool ShadowStackGCLowering::runOnFunction(Function &F) { 373 // Quick exit for functions that do not use the shadow stack GC. 374 if (!F.hasGC() || 375 F.getGC() != std::string("shadow-stack")) 376 return false; 377 378 LLVMContext &Context = F.getContext(); 379 380 // Find calls to llvm.gcroot. 381 CollectRoots(F); 382 383 // If there are no roots in this function, then there is no need to add a 384 // stack map entry for it. 385 if (Roots.empty()) 386 return false; 387 388 // Build the constant map and figure the type of the shadow stack entry. 389 Value *FrameMap = GetFrameMap(F); 390 Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F); 391 392 // Build the shadow stack entry at the very start of the function. 393 BasicBlock::iterator IP = F.getEntryBlock().begin(); 394 IRBuilder<> AtEntry(IP->getParent(), IP); 395 396 Instruction *StackEntry = 397 AtEntry.CreateAlloca(ConcreteStackEntryTy, nullptr, "gc_frame"); 398 399 while (isa<AllocaInst>(IP)) 400 ++IP; 401 AtEntry.SetInsertPoint(IP->getParent(), IP); 402 403 // Initialize the map pointer and load the current head of the shadow stack. 404 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead"); 405 Instruction *EntryMapPtr = 406 CreateGEP(Context, AtEntry, StackEntry, 0, 1, "gc_frame.map"); 407 AtEntry.CreateStore(FrameMap, EntryMapPtr); 408 409 // After all the allocas... 410 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 411 // For each root, find the corresponding slot in the aggregate... 412 Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root"); 413 414 // And use it in lieu of the alloca. 415 AllocaInst *OriginalAlloca = Roots[I].second; 416 SlotPtr->takeName(OriginalAlloca); 417 OriginalAlloca->replaceAllUsesWith(SlotPtr); 418 } 419 420 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't 421 // really necessary (the collector would never see the intermediate state at 422 // runtime), but it's nicer not to push the half-initialized entry onto the 423 // shadow stack. 424 while (isa<StoreInst>(IP)) 425 ++IP; 426 AtEntry.SetInsertPoint(IP->getParent(), IP); 427 428 // Push the entry onto the shadow stack. 429 Instruction *EntryNextPtr = 430 CreateGEP(Context, AtEntry, StackEntry, 0, 0, "gc_frame.next"); 431 Instruction *NewHeadVal = 432 CreateGEP(Context, AtEntry, StackEntry, 0, "gc_newhead"); 433 AtEntry.CreateStore(CurrentHead, EntryNextPtr); 434 AtEntry.CreateStore(NewHeadVal, Head); 435 436 // For each instruction that escapes... 437 EscapeEnumerator EE(F, "gc_cleanup"); 438 while (IRBuilder<> *AtExit = EE.Next()) { 439 // Pop the entry from the shadow stack. Don't reuse CurrentHead from 440 // AtEntry, since that would make the value live for the entire function. 441 Instruction *EntryNextPtr2 = 442 CreateGEP(Context, *AtExit, StackEntry, 0, 0, "gc_frame.next"); 443 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead"); 444 AtExit->CreateStore(SavedHead, Head); 445 } 446 447 // Delete the original allocas (which are no longer used) and the intrinsic 448 // calls (which are no longer valid). Doing this last avoids invalidating 449 // iterators. 450 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 451 Roots[I].first->eraseFromParent(); 452 Roots[I].second->eraseFromParent(); 453 } 454 455 Roots.clear(); 456 return true; 457 } 458